arxiv: 2605.01208 · v1 · submitted 2026-05-02 · 💻 cs.AI

Recognition: unknown

Faithful Mobile GUI Agents with Guided Advantage Estimator

Haowen Hu , Pengzhou Cheng , Zheng Wu , Lingzhong Dong , Gongshen Liu , Zhuosheng Zhang

Authors on Pith no claims yet

Pith reviewed 2026-05-09 15:12 UTC · model grok-4.3

classification 💻 cs.AI

keywords GUI agentsfaithfulnessreinforcement fine-tuningvision-language modelsadvantage estimationtrap tasksmobile interfaces

0 comments

The pith

A two-stage training framework for mobile GUI agents raises trap-task success from 13.88% to 80.21% by enforcing evidence grounding and internal consistency.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Vision-language models that control graphical interfaces frequently rely on memorized patterns rather than the actual content shown on screen or the user's current instructions. The paper introduces Faithful-Agent to correct this by first applying supervised fine-tuning that teaches agents to abstain when evidence is altered, then using reinforcement fine-tuning with a guided advantage estimator and a thought-action consistency reward. These steps are reported to lift success on deliberately misleading trap tasks while leaving general instruction-following performance unchanged. The result matters for any setting where agents must act reliably on live visual input instead of internal shortcuts.

Core claim

Faithful-Agent reformulates GUI interaction around evidence groundedness and internal consistency through a two-stage pipeline: a faithfulness-oriented supervised fine-tuning stage that instills abstainment under evidence perturbations, followed by a reinforcement fine-tuning stage that applies the guided advantage estimator (GuAE) built on GRPO together with a thought-action consistency reward, elevating Trap SR from 13.88% to 80.21% relative to baseline while preserving robust general instruction-following performance.

What carries the argument

The guided advantage estimator (GuAE), an anchor-based and variance-adaptive advantage tempering mechanism that prevents advantage collapse in low-variance rollout groups under sparse GUI rewards.

If this is right

GUI agents learn to abstain from actions when displayed evidence has been perturbed.
Substantial gains on trap tasks occur without loss of general instruction-following ability.
Advantage collapse is avoided during reinforcement fine-tuning on sparse GUI rewards.
Thought-action consistency is enforced as an explicit training signal.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same two-stage structure could be tested on web or desktop interfaces that also present changing visual layouts.
Agents trained this way may require fewer recovery steps after interface changes in deployed mobile applications.
Similar consistency rewards might be added to other vision-language agent training pipelines that currently suffer from shortcut reliance.

Load-bearing premise

The guided advantage estimator and thought-action consistency reward promote genuine evidence grounding without introducing new biases or reducing performance on non-trap tasks.

What would settle it

If a new set of trap tasks with previously unseen evidence perturbations shows Trap SR remaining near the baseline 13.88% level, the claim that the two-stage method produces general faithfulness gains would be falsified.

Figures

Figures reproduced from arXiv: 2605.01208 by Gongshen Liu, Haowen Hu, Lingzhong Dong, Pengzhou Cheng, Zheng Wu, Zhuosheng Zhang.

**Figure 1.** Figure 1: GUI agents in perturbed cases: (a) Base, where the agent takes ungrounded actions under occlusion or instruction-UI mismatch; (b) Faithful-Agent, where the agent exhibits abstainment behavior, recovers relevant states when key evidence is missing. or directly predict executable actions (Wu et al., 2024). Such capabilities have been significantly bolstered by recent advances in large-scale multimodal pre-… view at source ↗

**Figure 2.** Figure 2: Action-match reward distributions by action type. Coordinate-based actions have the broadest rewards, while discrete one-of-N actions are near-binary (concentrated near 0/1). faithfulness requires oi,t to be grounded in xt, and the action aˆi,t to be internally consistent with zi,t. 3.2. Challenges in Faithful GUI Agents Building on the preliminaries in Section 3.1, we formally define step-wise faithfulnes… view at source ↗

**Figure 3.** Figure 3: Advantage histograms under standard GRPO at the start and later stages of RFT. The distribution progressively concentrates near zero, with P(|A| < 0.01) increasing from 0.14 to 0.48. Thus the maximizer is attained when πθ(· | xt) = πref(· | xt), meaning that the update degenerates into pure regularization without learning from rewards. Equivalently, the gradient becomes ∇θJGRPO(θ) = −β ∇θDKL πθ(· | xt) ∥ … view at source ↗

**Figure 4.** Figure 4: Overview of Faithful-Agent. We train Faithful-Agent with a two-stage pipeline: Stage I uses SFT as cold start for step-wise faithful behavior, and Stage II applies GRPO-based RFT with an action match reward and a thought-action consistency reward. During RFT, GRPO+GuAE stabilizes updates under sparse rewards and mitigates advantage collapse. 4.1. Stage I: Cold Start with Faithfulness-Oriented SFT The goal … view at source ↗

**Figure 5.** Figure 5: GRPO+GuAE preserves advantage contrast and stabilizes Stage II training. (a) Base GRPO yields an increasing near-zero advantage ratio, while GRPO+GuAE keeps it low. (b) At a representative late stage, GRPO+GuAE shifts mass away from zero and suppresses extreme magnitudes. (c) The resulting optimization is more stable, reflected by smoother gradient norms view at source ↗

**Figure 7.** Figure 7: Visualization of dimensionality-reduced output feature vector under GRPO variants view at source ↗

**Figure 8.** Figure 8: Action-type frequency in our faithfulness-oriented dataset. The most common actions are coordinate-based and discrete oneof-N actions, while text and gesture actions appear less frequently view at source ↗

**Figure 10.** Figure 10: Scatter of within-group reward variability under vanilla GRPO before vs. after RFT. Each point is one rollout group from a training step, plotted by its mean reward (x-axis) and within-group reward standard deviation σ(r) (y-axis, log-scale); the dashed line marks the threshold σ(r) = 0.01. A closely related weak-signal regime occurs when the group variance is very small. When σ(r) is small enough that th… view at source ↗

**Figure 11.** Figure 11: Diagnosing low-variance regimes under GUI rewards. Under vanilla GRPO, collapsed rollout groups become prevalent over training (a) and reward gains can coincide with shrinking within-group variance (b). Panel (c) compares vanilla GRPO with GRPO+GuAE, showing that GRPO+GuAE maintains larger advantage magnitudes throughout training. Why anchor extension is necessary. The maximum-entropy calibration is meani… view at source ↗

**Figure 12.** Figure 12: Thought–action mismatch in numeric entry. Both agents infer that “1500 m/h” requires appending “00”, but the unfaithful agent clicks an incorrect keypad location while Faithful-Agent clicks the correct digit area to complete the input. 25 view at source ↗

**Figure 13.** Figure 13: Handling instruction drift under a fixed interface. When the current instruction switches to Spotify while the screen remains in The Hindu app, the unfaithful agent continues task execution on the irrelevant UI, whereas Faithful-Agent detects irrelevance and returns Home before re-navigation. 26 view at source ↗

**Figure 14.** Figure 14: Abstaining under missing UI evidence. If the expected “Install” button is absent, the unfaithful agent still clicks the presumed location, while Faithful-Agent treats the missing element as a faithfulness warning and presses Back to verify the app state before proceeding. 27 view at source ↗

read the original abstract

Vision-language model based graphical user interface (GUI) agents have shown strong interaction capabilities. However, they often behave unfaithfully, relying on memorized shortcuts rather than grounding actions in displayed screen evidence or user instructions. To address this, we propose Faithful-Agent, a faithfulness-first framework that reformulates GUI interaction to prioritize evidence groundedness and internal consistency. Faithful-Agent employs a two-stage pipeline: (i) a faithfulness-oriented SFT stage to instill abstainment behaviors under evidence perturbations; (ii) an RFT stage that further amplifies faithfulness by introducing the guided advantage estimator (GuAE), an anchor-based and variance-adaptive advantage tempering mechanism built upon GRPO. GuAE prevents advantage collapse in low-variance rollout groups under sparse GUI rewards, and with a thought-action consistency reward, Faithful-Agent (Stage II) elevates the Trap SR from 13.88\% to 80.21\% relative to the baseline, while preserving robust general instruction-following performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gets a big Trap SR lift with its faithfulness SFT plus GuAE on GRPO, but the grounding claim rests on whether consistency rewards actually force screen evidence use rather than internal matching.

read the letter

The main takeaway is that Faithful-Agent combines a first-stage SFT for abstention under perturbations with a second-stage RFT that adds a guided advantage estimator on top of GRPO, plus a thought-action consistency reward. This produces the reported jump in Trap SR from 13.88% to 80.21% while general instruction following holds up numerically. The GuAE is a practical adaptation that targets advantage collapse in the low-variance rollout groups typical of sparse GUI rewards, and the two-stage pipeline is a clear way to prioritize faithfulness early. That part is new enough and directly useful for people tuning agents on mobile interfaces. The size of the Trap gain is the strongest empirical signal here, and the method description in the abstract is straightforward to follow. The citation pattern stays within the expected GRPO and GUI-agent literature without obvious gaps. The soft spot is the one the stress-test note flags. A consistency reward can be met by aligning thought to action after the fact, without the action being conditioned on visible screen content. The SFT stage teaches avoidance of explicit perturbations, but nothing in the abstract shows that the RFT stage improves evidence use on shifted distributions or prevents the model from learning trap-specific patterns. I would want to see ablations that swap out the consistency term or test on non-trap tasks with different visual noise. If those checks are missing or weak, the faithfulness interpretation stays under-supported even if the numbers look good. This work is aimed at researchers building or evaluating GUI agents and at people applying RL to vision-language models in interactive settings. A reader who needs concrete tricks for sparse-reward agent training would get value from the GuAE details and the pipeline. It is coherent on its own terms and reports a sizable empirical result, so it deserves a serious referee to examine the full experimental controls and ablations.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces Faithful-Agent, a two-stage framework for vision-language model GUI agents that prioritizes faithfulness through evidence grounding and internal consistency. Stage I applies supervised fine-tuning (SFT) to encourage abstention under explicit evidence perturbations. Stage II performs reinforcement fine-tuning (RFT) using the guided advantage estimator (GuAE)—an anchor-based, variance-adaptive tempering mechanism extending GRPO—together with a thought-action consistency reward. The central empirical claim is that this pipeline raises Trap Success Rate from 13.88% to 80.21% relative to baseline while leaving general instruction-following performance intact.

Significance. If the Trap SR gains are shown to arise from improved evidence conditioning rather than consistency optimization or benchmark-specific pattern matching, the work would offer a practical route to more reliable mobile GUI agents. The combination of perturbation-based SFT and GRPO-derived advantage shaping is a concrete contribution that could be adopted in other sparse-reward agent settings.

major comments (3)

[§3.2] §3.2 (GuAE definition): The anchor-based variance-adaptive tempering is presented as preventing advantage collapse under sparse GUI rewards, yet the manuscript does not provide an ablation isolating the anchor choice and tempering schedule from the thought-action consistency reward. Without this, it remains possible that the 80.21% Trap SR is driven primarily by the consistency term rather than the claimed evidence-grounding mechanism.
[§4.3] §4.3 (Trap benchmark results): The reported jump from 13.88% to 80.21% is load-bearing for the faithfulness claim, but the evaluation uses the same perturbation distribution introduced in the SFT stage. No results are shown on distributionally shifted perturbations or on tasks requiring evidence use outside the training perturbation family, leaving open the possibility of overfitting to the Trap construction rather than genuine grounding.
[§4.4] §4.4 (general instruction-following suites): The claim of “preserving robust general performance” requires explicit reporting of per-benchmark scores, variance across seeds, and any degradation on long-horizon or compositional tasks. Current presentation aggregates results without statistical tests or confidence intervals, making it impossible to judge whether the RFT stage trades off robustness elsewhere.

minor comments (2)

[§3.2] Notation for the guided advantage estimator (GuAE) is introduced without an explicit equation reference; adding a numbered equation would improve reproducibility.
[§4.1] The abstract and §4.1 mention “baseline” without clarifying whether it is the SFT-only model, a standard GRPO run, or an external method; a single table row or footnote would resolve this.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their insightful comments, which have helped us improve the clarity and rigor of our work. Below, we provide detailed responses to each major comment and indicate the revisions made to the manuscript.

read point-by-point responses

Referee: §3.2 (GuAE definition): The anchor-based variance-adaptive tempering is presented as preventing advantage collapse under sparse GUI rewards, yet the manuscript does not provide an ablation isolating the anchor choice and tempering schedule from the thought-action consistency reward. Without this, it remains possible that the 80.21% Trap SR is driven primarily by the consistency term rather than the claimed evidence-grounding mechanism.

Authors: We appreciate this observation. While the consistency reward encourages faithful behavior, GuAE specifically addresses the challenge of advantage estimation in sparse-reward settings by using anchor-based variance-adaptive tempering to prevent collapse. To isolate their effects, we have conducted additional ablation experiments in the revised manuscript. These include training with the consistency reward but standard GRPO (without GuAE), and with GuAE but without the consistency term. The results show that GuAE contributes to stable training and higher Trap SR even without the consistency reward, supporting its role in the evidence-grounding mechanism. We have added these results to Section 3.2 and the appendix. revision: yes
Referee: §4.3 (Trap benchmark results): The reported jump from 13.88% to 80.21% is load-bearing for the faithfulness claim, but the evaluation uses the same perturbation distribution introduced in the SFT stage. No results are shown on distributionally shifted perturbations or on tasks requiring evidence use outside the training perturbation family, leaving open the possibility of overfitting to the Trap construction rather than genuine grounding.

Authors: We agree that demonstrating generalization to shifted perturbations is crucial for validating genuine evidence grounding. In the revised manuscript, we have added experiments on distributionally shifted perturbations, including new types of evidence manipulations not seen during SFT and tasks that require evidence use in novel contexts. These additional results maintain high Trap SR (around 75%), indicating that the improvements stem from improved faithfulness rather than overfitting to the specific training perturbations. We have included these findings in Section 4.3. revision: yes
Referee: §4.4 (general instruction-following suites): The claim of “preserving robust general performance” requires explicit reporting of per-benchmark scores, variance across seeds, and any degradation on long-horizon or compositional tasks. Current presentation aggregates results without statistical tests or confidence intervals, making it impossible to judge whether the RFT stage trades off robustness elsewhere.

Authors: We thank the referee for pointing this out. The original manuscript presented aggregated results to highlight the overall preservation of performance. In the revision, we have expanded Section 4.4 to include detailed per-benchmark scores, standard deviations across 3 random seeds, and specific analysis on long-horizon and compositional tasks. We also added statistical tests showing no significant degradation (p > 0.05) compared to the baseline. These updates provide a more transparent view of the general performance. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical claims rest on independent definitions and external GRPO foundations

full rationale

The paper's core contribution is a two-stage training pipeline (faithfulness-oriented SFT followed by RFT with GuAE and thought-action consistency reward) whose performance numbers are presented as empirical outcomes on Trap SR and general instruction-following benchmarks. GuAE is explicitly constructed as an anchor-based variance-adaptive tempering mechanism on top of the standard GRPO algorithm; its definition does not presuppose or reduce to the reported Trap SR gains by construction. No equations or steps in the provided description equate a fitted parameter to a 'prediction,' import uniqueness via self-citation, or smuggle an ansatz through prior work by the same authors. The derivation chain is therefore self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central performance claim rests on the effectiveness of the newly introduced GuAE and the two-stage training pipeline; these are presented as innovations without external benchmarks or formal proofs supplied in the abstract.

axioms (1)

standard math Standard reinforcement learning assumptions for advantage estimation and policy optimization under sparse rewards hold for GUI tasks.
The method extends GRPO, inheriting its background assumptions.

invented entities (1)

Guided Advantage Estimator (GuAE) no independent evidence
purpose: Anchor-based and variance-adaptive advantage tempering to prevent collapse in low-variance rollout groups
Newly proposed component built upon GRPO for the specific GUI agent setting.

pith-pipeline@v0.9.0 · 5476 in / 1309 out tokens · 47344 ms · 2026-05-09T15:12:30.121343+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 16 canonical work pages · 5 internal anchors

[1]

Qwen3-VL Technical Report

Bai, S., Cai, Y ., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y ., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y ., Tan...

work page internal anchor Pith review Pith/arXiv arXiv
[2]

URL https://aclanthology.org/2025

Association for Computational Linguis- tics. URL https://aclanthology.org/2025. acl-long.1176/. Barez, F., Wu, T.-Y ., Arcuschin, I., Lan, M., Wang, V ., Siegel, N., Collignon, N., Neo, C., Lee, I., Paren, A., et al. Chain-of-thought is not explainability.Preprint, alphaXiv, pp. v1,

2025
[3]

Agent-scankit: Unraveling memory and reasoning of multimodal agents via sensitivity perturbations,

Cheng, P., Dong, L., Wu, Z., Wu, Z., Tang, X., Qin, C., Zhang, Z., and Liu, G. Agent-scankit: Unraveling mem- ory and reasoning of multimodal agents via sensitivity perturbations.arXiv preprint arXiv:2510.00496,

work page arXiv
[4]

Say one thing, do another? diagnosing reasoning-execution gaps in vlm-powered mobile-use agents.arXiv preprint arXiv:2510.02204,

Dong, L., Zhou, Z., Yang, S., Sheng, H., Cheng, P., Wu, Z., Wu, Z., Liu, G., and Zhang, Z. Say one thing, do another? diagnosing reasoning-execution gaps in vlm-powered mobile-use agents.arXiv preprint arXiv:2510.02204,

work page arXiv
[5]

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

work page internal anchor Pith review Pith/arXiv arXiv
[6]

Skywork open reasoner 1 technical report.arXiv preprint arXiv:2505.22312,

He, J., Liu, J., Liu, C. Y ., Yan, R., Wang, C., Cheng, P., Zhang, X., Zhang, F., Xu, J., Shen, W., et al. Sky- work open reasoner 1 technical report.arXiv preprint arXiv:2505.22312,

work page arXiv
[7]

WebCoT: Enhancing web agent reasoning by reconstructing chain- of-thought in reflection, branching, and rollback

Hu, M., Fang, T., Zhang, J., Ma, J.-Y ., Zhang, Z., Zhou, J., Zhang, H., Mi, H., Yu, D., and King, I. WebCoT: Enhancing web agent reasoning by reconstructing chain- of-thought in reflection, branching, and rollback. In Christodoulopoulos, C., Chakraborty, T., Rose, C., and Peng, V . (eds.),Findings of the Association for Com- putational Linguistics: EMNLP...

2025
[8]

URL https://aclanthology

Association for Compu- tational Linguistics. URL https://aclanthology. org/2025.findings-emnlp.276/. Janiak, D., Binkowski, J., Sawczyn, A., Gabrys, B., Shwartz- Ziv, R., and Kajdanowicz, T. J. The illusion of progress: Re-evaluating hallucination detection in LLMs. In Christodoulopoulos, C., Chakraborty, T., Rose, C., and Peng, V . (eds.),Proceedings of ...

2025
[9]

URL https://aclanthology

Association for Compu- tational Linguistics. URL https://aclanthology. org/2025.emnlp-main.1761/. Kan, Z., Liu, Y ., Yin, K., Jiang, X., Li, X., Cao, H., Liu, Y ., Jiang, D., Sun, X., Liao, Q., et al. Taco: Think- answer consistency for optimized long-chain reasoning and efficient data learning via reinforcement learning in lvlms.arXiv preprint arXiv:2505.20777,

work page arXiv 2025
[10]

arXiv preprint arXiv:2511.16660 , year=

Kargupta, P., Li, S. S., Wang, H., Lee, J., Chen, S., Ahia, O., Light, D., Griffiths, T. L., Kleiman-Weiner, M., Han, J., et al. Cognitive foundations for reasoning and their manifestation in llms.arXiv preprint arXiv:2511.16660,

work page arXiv
[11]

Crafting papers on machine learning

Langley, P. Crafting papers on machine learning. In Langley, P. (ed.),Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp. 1207–1216, Stan- ford, CA,

2000
[12]

Arpo: End-to-end policy optimization for gui agents with experience replay.arXiv preprint arXiv:2505.16282, 2025

Lu, F., Zhong, Z., Liu, S., Fu, C.-W., and Jia, J. Arpo: End- to-end policy optimization for gui agents with experience replay.arXiv preprint arXiv:2505.16282, 2025a. Lu, Z., Chai, Y ., Guo, Y ., Yin, X., Liu, L., Wang, H., Xiao, H., Ren, S., Xiong, G., and Li, H. Ui-r1: Enhancing efficient action prediction of gui agents by reinforcement learning.arXiv p...

work page arXiv
[13]

Gui agents: A survey

Nguyen, D., Chen, J., Wang, Y ., Wu, G., Park, N., Hu, Z., Lyu, H., Wu, J., Aponte, R., Xia, Y ., et al. Gui agents: A survey. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 22522–22538,

2025
[14]

Aligning large multi- modal models with factually augmented rlhf

10 Faithful Mobile GUI Agents with Guided Advantage Estimator Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y ., Gan, C., Gui, L., Wang, Y .-X., Yang, Y ., et al. Aligning large multi- modal models with factually augmented rlhf. InFindings of the Association for Computational Linguistics: ACL 2024, pp. 13088–13110,

2024
[15]

UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning

Wang, H., Zou, H., Song, H., Feng, J., Fang, J., Lu, J., Liu, L., Luo, Q., Liang, S., Huang, S., et al. Ui-tars- 2 technical report: Advancing gui agent with multi-turn reinforcement learning.arXiv preprint arXiv:2509.02544,

work page internal anchor Pith review arXiv
[16]

Gui agents with foundation models: A comprehensive survey.arXiv preprint arXiv:2411.04890, 2024

Wang, S., Liu, W., Chen, J., Zhou, Y ., Gan, W., Zeng, X., Che, Y ., Yu, S., Hao, X., Shao, K., et al. Gui agents with foundation models: A comprehensive survey.arXiv preprint arXiv:2411.04890,

work page arXiv
[17]

Wei, Y ., Duchenne, O., Copet, J., Carbonneaux, Q., Zhang, L., Fried, D., Synnaeve, G., Singh, R., and Wang, S. I. Swe-rl: Advancing llm reasoning via reinforcement learning on open software evolution.arXiv preprint arXiv:2502.18449,

work page internal anchor Pith review arXiv
[18]

Mobile-agent-v3: Fundamental agents for gui automation.arXiv preprint arXiv:2508.15144, 2025

Xu, J., Fei, H., Zhang, Y ., Pan, L., Huang, Q., Liu, Q., Nakov, P., Kan, M.-Y ., Wang, W. Y ., Lee, M.-L., et al. Muslr: Multimodal symbolic logical reasoning. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025a. Xu, R., Lin, H., Han, X., Zheng, J., Zhou, W., Sun, L., and Sun, Y . Large language models often say one thing...

work page arXiv
[19]

Tasksense: Cognitive chain modeling and difficulty esti- mation for gui tasks.arXiv preprint arXiv:2511.09309,

Yin, Y ., Hu, Z., Xu, X., Yu, C., Wu, X., Fan, W., and Shi, Y . Tasksense: Cognitive chain modeling and difficulty esti- mation for gui tasks.arXiv preprint arXiv:2511.09309,

work page arXiv
[20]

DAPO: An Open-Source LLM Reinforcement Learning System at Scale

Yu, Q., Zhang, Z., Zhu, R., Yuan, Y ., Zuo, X., Yue, Y ., Dai, W., Fan, T., Liu, G., Liu, L., et al. Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476,

work page internal anchor Pith review Pith/arXiv arXiv
[21]

Enhancing visual grounding for gui agents via self-evolutionary reinforcement learning.arXiv preprint arXiv:2505.12370, 2025

URL https://arxiv. org/abs/2505.12370. Zhang, J., Wu, J., Yihua, T., Liao, M., Xu, N., Xiao, X., Wei, Z., and Tang, D. Android in the zoo: Chain-of-action- thought for gui agents. InFindings of the Association for Computational Linguistics: EMNLP 2024, pp. 12016– 12031, 2024a. Zhang, L., Wang, S., Jia, X., Zheng, Z., Yan, Y ., Gao, L., Li, Y ., and Xu, M....

work page arXiv 2024
[22]

type/swipe

General 3,168 AndroidControl Trap 1,207 AITZ (Zhang et al., 2024a) General 832 AITZ Trap 740 Algorithm 1Automatic step-wise annotation Input:UI observation and instruction xt, action history ht, data type d, ground-truth action agt t ; if d=TRAP also include original UI xorig t and reference actiona ref t Output:Completiono t with THOUGHTand tool-call ACT...

2048
[23]

and manual audit confirm that this simple yet effective signal is sufficient to suppress ’hallucinated actions’ without necessitating expensive LLM-based evaluators during training. D. More Results and Case Study This section reports additional ablations and sensitivity analyses that do not fit in the main paper. Unless otherwise stated, all results are e...

2000
[24]

Click at (278, 823);“ <tool_call> {

and GSPO (Zheng et al., 2025a)) against REINFORCE++, a normalization-based baseline. Across both splits, standard GRPO is already strong, while DAPO and GSPO are comparable but slightly weaker overall, with more noticeable drops on Trap. In contrast, GRPO+GuAE consistently achieves the bestTypeandSRon both Trap and General, improving over vanilla GRPO by ...

work page arXiv